We present TWIST, a novel self-supervised representation learning method by classifying large-scale unlabeled datasets in an end-to-end way. We employ a siamese network terminated by a softmax operation to produce twin class distributions of two augmented images. Without supervision, we enforce the class distributions of different augmentations to be consistent. In the meantime, we regularize the class distributions to make them sharp and diverse. Specifically, we minimize the entropy of the distribution for each sample to make the class prediction for each sample assertive and maximize the entropy of the mean distribution to make the predictions of different samples diverse. In this way, TWIST can naturally avoid the trivial solutions without specific designs such as asymmetric network, stop-gradient operation, or momentum encoder. Different from the clustering-based methods which alternate between clustering and learning, our method is a single learning process guided by a unified loss function. As a result, TWIST outperforms state-of-the-art methods on a wide range of tasks, including unsupervised classification, linear classification, semi-supervised learning, transfer learning, and some dense prediction tasks such as detection and segmentation.
View on arXiv